Okay so I do have to apply the same rejection. There was a misunderstanding because ft can reuse un/mixing matrices but the rejection of components is a separate step which "backprojects an ICA (or similar) decomposition to the channel level after removing the independent components that contain the artifacts." So what I did was simply decomposing without rejecting anything.
The empty room noise covariance matrix does change a lot if I keep the same rejection as in the subject recording.